How to Run an AI Visibility Audit

An AI visibility audit measures how often AI engines mention, cite and recommend your brand for a fixed set of buyer prompts, finds out why, and ends with a ranked fix list. Over five working days you set the scope, run 30-50 prompts across four engines, score the answers, map the sources engines rely on, check crawler access and page quality, and present a baseline plus actions.

What you need: a spreadsheet, accounts for ChatGPT, Perplexity and Gemini, access to Google (for AI Overviews and AI Mode), robots.txt and log access, and an hour with sales.

On this page

• Day 1: How do you set the scope and prompts?
• Day 2: How do you run the scans?
• Day 3: How do you score and map sources?
• From MentionX data: what audit-scale scans show
• Day 4: Are there access or page problems?
• Worked example: Ledgerline's audit findings
• Day 5: How do you prioritise and present?
• How much time does each day really take?
• What should you check in the logs?
• What does a finished audit include?
• How do you present findings so they get acted on?
• How do you audit for a client?
• How do you handle multiple markets or languages?
• How often should you re-audit?
• What are the common mistakes?
• How is this different from an SEO audit?
• Common questions
• Sources

Engine and crawler documentation checked on 30 October 2026. "Our view" marks MentionX opinion. Ledgerline is fictional.

Day 1: How do you set the scope and prompts?

Scope decisions (write them down):

Prompt set (30-50): collect real wording from:

• Sales call notes ("prospects ask whether...")
• Support tickets and onboarding questions
• Reddit and community threads in your category
• Review-site questions and comparison searches
• Search Console queries phrased as questions

Group by intent: category, use case, comparison, pricing. Leave brand-name prompts in a separate tab.

Day 2: How do you run the scans?

For each prompt × engine × run, record:

• Date, engine, mode
• Brands mentioned, in order
• Whether each brand was recommended for the stated need
• All cited URLs
• Notes on wrong facts about you

Use fresh sessions and the same location throughout. If Google shows no AI Overview, record "not triggered"; don't count it as a miss.

Copy the full answer text into a column or a linked document. You'll need it to check claims later.

Day 3: How do you score and map sources?

Score per engine and overall:

• Mention rate, recommendation rate, citation rate
• Share of voice vs competitors (method)
• Optional single score (method)

Map sources: extract domains from every cited URL, count them, and rank. For the top 20, mark:

Illustrative row.

This table is usually the most useful page in the audit.

From MentionX data: what audit-scale scans show

Method. Read-only aggregate queries over MentionX scans from 22 August to 29 October 2026 (UTC): 54 scans and 2,217 engine runs (2,154 completed) across 18 anonymised projects, mostly B2B software, and six engines called through APIs; 9,750 cited URLs. Perplexity, Copilot and DeepSeek figures come from a single day (26 October), so treat them as a snapshot.

Limits. A small, non-random sample. API answers can differ from consumer apps, which a manual audit uses. Page types come from URL patterns, so many URLs stay unclassified. We have no crawler log data in this sample, so the access checks below rely on your own logs.

Our view: blog posts and homepages make up the largest identifiable share of cited pages here. In most audits, the biggest wins are a few existing pages made clearer, plus listings on the third-party pages engines already cite.

Day 4: Are there access or page problems?

Access checks:

1. robots.txt groups for OAI-SearchBot, GPTBot, PerplexityBot, Claude-SearchBot, ClaudeBot, Google-Extended, Googlebot, Bingbot (guide).
2. CDN/WAF bot rules.
3. Server logs: do these agents get 200s on key pages?
4. Rendering: is key content in the HTML?
5. Indexing in Google and Bing for your most relevant pages.

Page checks for the 5-10 pages closest to your priority prompts:

• Answer in the first 40-60 words?
• Question-shaped headings matching follow-ups?
• Comparison table, dated facts, named author?
• Snippet controls (nosnippet, max-snippet) not limiting quotes (Google)?

Accuracy checks: list every wrong fact found in answers, and find where it likely comes from (an old pricing page, an outdated review, a directory listing).

Worked example: Ledgerline's audit findings

Ledgerline is a fictional accounts-payable tool. Figures are illustrative. The audit covered 36 unbranded prompts, four engines and two runs: 288 runs, of which 281 completed.

Score: impact minus effort, ties broken by impact. The CDN rule takes ten minutes to fix and unblocks a whole engine, so it goes first. llms.txt goes last: it's cheap, but nothing in the answers points to it.

Day 5: How do you prioritise and present?

Score each finding on impact and effort (1-3 each). Typical top of the list:

1. Access blockers (high impact, low effort).
2. Wrong facts with a clear source (high impact, low effort).
3. Missing from the top 5 cited third-party domains (high impact, medium effort).
4. Rewriting priority pages answer-first (medium-high impact, medium effort).
5. New pages for topics where no page of yours fits (medium impact, medium effort).

Present in 8 slides:

1. Scope and method
2. Headline: mention, recommendation, citation rates; SOV
3. By engine
4. By topic
5. Source map
6. Access and page findings
7. Accuracy issues
8. Fix list with owners and dates

Attach the raw log.

How much time does each day really take?

For one brand, 40 prompts, four engines and two runs, a realistic budget looks like this:

That's roughly 20 hours. The scan day is the easiest to shorten with a tool. The source map and accuracy tracing are the hardest to shorten, and they produce most of the fixes.

What should you check in the logs?

If you have server or CDN logs, pull 30 days and filter by the user agents in your robots.txt review. For each agent, look at three things: whether it reaches your priority pages at all, the share of 4xx and 5xx responses, and whether requests come from the IP ranges the operator publishes. Google explains how to verify its crawlers, and Perplexity publishes its IP list. A user agent string on its own can be faked, so verify before you treat traffic as a real crawler.

What does a finished audit include?

• Scope doc with competitor set, engines and prompt version
• Raw answer log
• Rates and SOV per engine and topic
• Top-20 cited domains with presence flags
• Access findings with evidence (log lines, robots.txt excerpts)
• Page scores for priority pages
• Accuracy issues with likely sources
• Ranked fix list
• Re-audit date

How do you present findings so they get acted on?

Audits stall when findings are true but nobody owns them. A few habits help:

• One owner per fix, by name, with a date. "Content team" is not an owner.
• Evidence on the slide. A log line, an answer quote and the cited URL persuade faster than a rate.
• Separate quick wins. Put fixes under a day's work on their own slide; they build credibility for the bigger asks.
• Say what you won't do. If a channel or engine doesn't matter for your buyers, say so and move on.
• Book the re-scan date in the meeting, so results get checked.

How do you audit for a client?

Agree the competitor set and prompt list with the client before running anything, because disputes later usually come from scope. Share the raw answer log with the summary so they can check your work. Keep the client's brand and competitor names out of any shared templates or case studies unless they've agreed in writing.

How do you handle multiple markets or languages?

Run a separate prompt set per market, written by someone who speaks the language natively, not translated word for word. Buyers in Germany may ask about DATEV integration or GDPR hosting in a way US buyers never do. Report each market as its own baseline; averaging them hides the market where you're invisible.

How often should you re-audit?

Re-run the scans monthly (same prompts) and the full audit quarterly. Refresh the prompt set twice a year with new sales and support questions, and start a new trend line when you do.

What are the common mistakes?

• Letting marketing write the prompts without sales or support input.
• One run per prompt.
• Comparing engines without noting mode (e.g. ChatGPT with and without search).
• Reporting numbers without the raw answers behind them.
• Ending with insights and no owners.

How is this different from an SEO audit?

An SEO audit starts with your site; this audit starts with the answers. Many findings are off-site (review sites, forums, lists), and the main metric is presence in answers, not rankings. The two should be run together and share a fix list. See AEO vs SEO vs GEO.

Common questions

How long does an AI visibility audit take? About a week part-time for one brand, 30-50 prompts and four engines. Most of the time goes into running prompts and mapping sources.

What's the output of the audit? A baseline of mention, citation and recommendation rates and share of voice; a map of the sources engines cite; access and page findings; and a ranked fix list with owners.

Do I need server logs? They're the only way to confirm crawlers reach your pages. If you can't get them, note it as a limit and check robots.txt and CDN rules instead.

Should branded prompts be in the audit? Keep them in a separate tab. They're useful for checking accuracy, because engines almost always mention you when asked by name, but they inflate headline rates.

What if an engine gives no citations? Record the answer and the brands anyway. You can still score mentions and recommendations; you just can't map sources for that engine, which is worth saying in the method notes.

Who should run it? Someone who knows the category and can read answers critically, with help from sales for prompts and from engineering for logs.

When should I re-run the full audit? Quarterly, with monthly scans of the same prompts in between.

Can a tool do the audit? A tool can run the scans and count citations. Reading answers, tracing wrong facts and ranking fixes still need a person.

Sources

• Google Search Central, AI features and your website
• OpenAI, crawlers; Perplexity, crawlers; Anthropic, crawler FAQ
• Google Search Central, Introduction to robots.txt; Verifying Google requests
• Apache HTTP Server, Log files
• Microsoft, How web search works in Copilot
• NIST/SEMATECH, Confidence intervals for a proportion